[
https://issues.apache.org/jira/browse/YARN-2113?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15951445#comment-15951445
]
Jason Lowe commented on YARN-2113:
----------------------------------
bq. once user submit app3 with highest priority, it should get resource from
other apps if user3's UL doesn't meet.
I do not agree. The problem here is that the other users _are not above their
MULP_. I thought we were all in agreement that users who stay below their MULP
are not going to be preempted in-queue. If they are, then how do they protect
themselves from users who decide to jack up their priorities in arbitrary ways?
The only way is if they, too, participate in the priority arms race.
Let's think about how this is going to play out in practice. The default queue
config is 100% MULP. Alice shows up and submits a variety of apps that overall
consume 30% of the queue. Then Bob comes along later and submits a single,
large app that's big enough to saturate the queue. Bob happens to submit that
app with a slightly higher priority. If the proposed behavior is in place,
Bob's app will decimate *all* of Alice's active apps. Is that really desired?
What do admins tell users that are worried about getting their stuff preempted?
How does a user tailor their load so they won't be a victim of preemption?
What guarantee can they rely on?
> Add cross-user preemption within CapacityScheduler's leaf-queue
> ---------------------------------------------------------------
>
> Key: YARN-2113
> URL: https://issues.apache.org/jira/browse/YARN-2113
> Project: Hadoop YARN
> Issue Type: Sub-task
> Components: scheduler
> Reporter: Vinod Kumar Vavilapalli
> Assignee: Sunil G
> Attachments:
> TestNoIntraQueuePreemptionIfBelowUserLimitAndDifferentPrioritiesWithExtraUsers.txt,
> YARN-2113.0001.patch, YARN-2113.0002.patch, YARN-2113.0003.patch,
> YARN-2113.0004.patch, YARN-2113.v0.patch
>
>
> Preemption today only works across queues and moves around resources across
> queues per demand and usage. We should also have user-level preemption within
> a queue, to balance capacity across users in a predictable manner.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
---------------------------------------------------------------------
To unsubscribe, e-mail: yarn-issues-unsubscr...@hadoop.apache.org
For additional commands, e-mail: yarn-issues-h...@hadoop.apache.org
